Models

Models can be changed by adding a MODELS section to the config. The default model is Seq2Seq.

Informer

MODEL:
  NAME: "Informer"
  # Number of hidden units
  NUM_H_FEATS: 512
  # Number of encoders
  NUM_ENCODERS: 2
  # Number of decoders
  NUM_DECODERS: 1
  # Number of heads of self attention
  NUM_HEADS: 8
  # Smaller value leads to higher memory efficiency
  CONTRACTION_FACTOR: 5
  # int(NUM_H_FEATS * FF_EXPANSION_RATE) is channel size of conv block after self attention
  EXPANSION_RATE: 4.0
  # Decoder input series length (last DECODER_IN_LENGTH values are used)
  DECODER_IN_LENGTH: 168
  # Dropout rate
  DROPOUT_RATE: 0.05

Reference

Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting

N-BEATS

MODEL:
  NAME: "NBeats"
  # Number of hidden units
  NUM_H_FEATS: 512
  # Depth of each block (set small value if dataset has high mean and variance)
  DEPTH: 4
  # Number of blocks
  STACK_SIZE: 30
  # Block type (option: {"identity", "trend"})
  BLOCK_TYPE: "identity"
  # Polynomial degree (used only if BLOCK_TYPE == "trend")
  DEGREE: 2

Reference

N-BEATS: Neural basis expansion analysis for interpretable time series forecasting

SCINet

MODEL:
  NAME: "SCINet"
  # Number of levels
  DEPTH: 3
  # Kernel size of conv modules
  KERNEL_SIZE: 5
  # Expansion rate of conv modules
  EXPANSION_RATE: 4.0
  # Dropout rate
  DROPOUT_RATE: 0.5

Reference

Time Series is a Special Sequence: Forecasting with Sample Convolution and Interaction

Seq2Seq

MODEL:
  NAME: "Seq2Seq"
  # Number of hidden units
  NUM_H_FEATS: 64
  # Number of encoders
  NUM_ENCODERS: 2
  # Number of decoders
  NUM_DECODERS: 1

Reference

Sequence to Sequence Learning with Neural Networks